Release dates, availability, pricing notes and Frontrumors adoption assessments for 10 model families. Recommendations are editorial judgments, not product guarantees.
Claude: Claude Opus 5.5
Anthropic · Release date 2026-09-22 · GA
Anthropic's flagship line: Fable 5.1 (Mythos-class, 1 Sep), Opus 5.5 (22 Sep), Sonnet 5.5 (28 Sep) and Haiku 5.5 (7 Oct) landed within five weeks.
Context window
1M tokens
Input / output per million
$4 / $20
Editorial recommendation
Adopt Now
Anthropic list prices: Opus 5.5 $4/$20, Sonnet 5.5 $2/$10, Fable 5.1 $10/$50, Haiku 5.5 from $0.10/$0.50 (higher above 100K-token prompts).
Opus 5.5 (22 September) is GA and fully available, and leads the Opus tier on agentic coding with 1M context at $4/$20. No successor is announced or rumored; Anthropic's flagship releases have come 12 to 73 days apart since November, so the next one is a low-confidence cycle estimate, not a dated event. Nothing about adopting today creates a second migration.
Adopt now. Opus 5.5 (22 September) has 1M context and leads the Opus tier on agentic coding, and no successor is announced or rumored; the next point release is a cycle estimate only. If you need the absolute ceiling, Fable 5.1 sits above it at $10/$50; otherwise Opus 5.5 at $4/$20 is the tier to standardise on today.
Overview
Claude is Anthropic's flagship family. The current tier leader is Opus 5.5 (22 September 2026, GA) at $4/$20 per million tokens with a 1M-token context window. Above it sits the Mythos-class Fable 5.1 (1 September, $10/$50); below it, Sonnet 5.5 (28 September) and Haiku 5.5 (7 October) fill out the line, four launches in five weeks.
What's new in this generation
The 5.5 generation is Anthropic's autumn stack: Fable 5.1 opened it on 1 September, Opus 5.5 followed on 22 September positioned as a step forward in agentic coding, Sonnet 5.5 landed 28 September and Haiku 5.5 on 7 October. Flagship releases since November 2025 (Opus 4.5, 4.6, 4.7, 4.8, Fable 5, Opus 5, Fable 5.1, Opus 5.5) have arrived 12 to 73 days apart, averaging about 43 days.
Features & capabilities
Opus 5.5's headline facts are its 1M-token context and tier-leading agentic coding results, the combination engineering teams are benchmarking coding agents against this quarter, including against GPT-6 Astra at 2.5x the price. Fable 5.1 is the maximum-capability option in the family; pay its premium only where Opus demonstrably falls short.
Pricing
Opus 5.5 lists at $4/$20 per million tokens (in/out). Sonnet 5.5 lists at $2/$10 and Fable 5.1 at $10/$50. Haiku 5.5 starts at $0.10/$0.50 for prompts up to 100K tokens and rises to $0.50/$2.50 above that. All figures are from Anthropic's pricing page, checked 27 October 2026.
Release history
2025-11-24: Claude Opus 4.5
2026-02-05: Claude Opus 4.6
2026-04-16: Claude Opus 4.7
2026-05-28: Claude Opus 4.8
2026-06-09: Claude Fable 5
2026-07-24: Claude Opus 5
2026-09-01: Claude Fable 5.1
What may come next
Next Opus / Fable point release · Nov 2026 – Q1 2027 (estimated from cycle) · Low, cycle estimate only
GPT-6 arrived in stages: Astra (3 Sep, limited orgs first), Sol + Luna (22 Sep), the balanced GPT-6.1 Sol (29 Sep) and, on 26 Oct, GPT-6.1 Terra, a $1/$5 computer-use and vision tier.
Context window
1.05M tokens
Input / output per million
$2 / $10
Editorial recommendation
Adopt Now
OpenAI list prices: GPT-6.1 Sol $2/$10, GPT-6.1 Terra $1/$5, GPT-6 Astra $10/$50, GPT-6 Luna $0.10/$0.50. GPT-5.6 Sol promotional pricing runs at least through 21 Nov 2026; OpenAI has announced no end date.
GPT-6.1 Sol is available in paid Work and Codex plans and the API, according to OpenAI's launch page. Evaluate this balanced tier on current workload results; the separate Astra-tier point-refresh rumor has no announced date.
Adopt Sol where its balanced capability and $2/$10 pricing fit. Terra adds a cheaper computer-use tier, while Astra remains the higher-priced reasoning option. A rumored Astra refresh is not a reason to treat the available Sol API as a limited release.
Overview
GPT is OpenAI's flagship family. GPT-6.1 Sol (29 September 2026, available in paid Work/Codex plans and the API) is the balanced tier; GPT-6 Astra (3 September) is the flagship at $10/$50, surfacing as GPT-6 Pro in ChatGPT Pro. Context is 1.05M tokens across the GPT-6 and GPT-6.1 tiers on the API.
What's new in this generation
GPT-6 arrived as a sequence, not a launch: Astra opened 3 September for limited organizations, Sol + Luna followed 22 September, and the balanced GPT-6.1 Sol landed 29 September. OpenAI's Sol announcement confirms access in paid Work and Codex plans and the API; earlier October rollout reports describe the launch sequence, not a current access restriction on Sol. GPT-6.1 Terra, a $1/$5 computer-use and vision tier, followed on 26 October. The recent cycle is fast, GPT-5.5 (April), GPT-5.6 family (July), GPT-6 (September).
Features & capabilities
Astra is the tier for the hardest reasoning and flagship agents, with frontier pricing. Check the access requirements for that tier separately from Sol. Sol ($2/$10) is the balanced middle for production workloads; Terra ($1/$5) trades some general capability for fast visual reasoning and computer use; Luna ($0.10/$0.50) is the budget tier. Plan migrations around access, not just capability.
Pricing
Per OpenAI's pricing page: Astra $10/$50, GPT-6.1 Sol $2/$10, GPT-6.1 Terra $1/$5, Luna $0.10/$0.50. Separately, GPT-5.6 Sol's promotional pricing is guaranteed only through at least 21 November 2026; no end date has been announced.
Carbon (28 Oct) is Google's newest frontier model, initially restricted to selected defenders and partners. Argon reached Stable GA on 13 October.
Context window
Unconfirmed in Google's announcement
Input / output per million
$2 / $10
Editorial recommendation
Rolling Out
Carbon's announced introductory price is $2/$10 per million tokens, then $4/$20; Google gives no introductory end date. Access is restricted. Separately, Argon rises from $2/$10 to $4/$20 on 1 Dec 2026.
Google announced Carbon on 28 October with restricted access for cyber defenders and trusted partners through Fairwind. It is an official model, but public API and consumer availability have no announced date. Argon remains the generally available Gemini 4 option.
Evaluate Carbon if your organization has access. Use Argon for generally available deployments, and budget for its December price increase. Do not plan around an unannounced Carbon public-release date or equate Carbon with the separate, unconfirmed Pro rumor.
Overview
Carbon follows Argon as the newest Gemini 4 model. Google announced it on 28 October and began a restricted rollout rather than a public launch. Argon remains available through Google's developer and consumer products.
What's new in this generation
Google emphasizes cybersecurity, coding and longer agent workflows. Availability starts with a subset of cyber defenders, expanding gradually to trusted partners through Fairwind; broader developer, business and consumer access is still forthcoming.
Features & capabilities
Artificial Analysis reports Carbon at 59 with high reasoning on Intelligence Index v5, checked 1 November. That independent composite does not mean it leads every task. Google's announcement does not specify a context limit, so we leave that specification unconfirmed here.
Pricing
Google announces Carbon at an introductory $2/$10 per million input/output tokens, followed by $4/$20, without a dated end to the introductory period. This is separate from Argon's confirmed 1 December step-up. Prices and access were checked against Google's pages on 1 November.
Release history
2026-02-19: Gemini 3.1 Pro
2026-05-19: Gemini 3.5 Flash
2026-09-02: Gemini 3.8 Flash
2026-09-30: Gemini 4 Argon (announced)
2026-10-13: Gemini 4 Argon (Stable GA)
What may come next
Gemini 4 Carbon wider access · No public date announced · Official restricted rollout; wider timing unannounced
Grok 4.7 is a new, larger base model at the same $2/$6 price as 4.6 with 500K context. Grok 5 is not expected until Q1 2027, so there is no reason to wait.
Adopt now. Grok 4.7 is a new, larger base model at the same $2/$6 price as 4.6 with 500K context, the value pick among Western frontier APIs this quarter. Grok 5 is not expected until Q1 2027, so there is no reason to wait.
Overview
Grok is xAI's flagship family. Grok 4.7 (21 September 2026, GA) is a new larger base model with a 500K-token context at $2/$6 per million tokens, unchanged from Grok 4.6 pricing. The path here ran Grok 4 (July 2025), 4.5 (February 2026) and 4.6 (June 2026).
What's new in this generation
The 4.7 story is a bigger base at last generation's price. Released a day before Claude Opus 5.5, it anchors xAI's line while all eyes are on Grok 5, which xAI says is training on Colossus 2 for a rumored Q1 2027 window.
Features & capabilities
A 500K context at $2/$6 makes 4.7 the value-frontier option in our Benchmarks: strong general capability without flagship pricing. Grok 5's rumored native multimodal input, including video temporal reasoning, is next-generation territory, not a 4.7 feature.
Pricing
$2/$6 per million tokens up to 200K tokens of context (xAI charges higher rates above that), unchanged from Grok 4.6. No pricing has been announced for Grok 5, and whether it holds the 4.x price line is unknown.
Spark 1.3 (2 Sep, 1M context) powers the Muse agent. Wonder (muse-wonder-1.0) entered preview on 12 October; the Benchmarks page distinguishes Wonder from Wonder Full and dates each AA measurement.
Context window
1M tokens
Input / output per million
$1.25 / $4.25
Editorial recommendation
Wait
Spark 1.3 pricing. We could not verify a first-party price for Muse Wonder; OpenRouter lists $4/$20.
Pricing last checked: not verified for Muse Spark 1.3
Spark 1.3 is a solid GA model, but Muse Wonder, Meta's most capable model, has been in preview since 12 October and is expected to reach GA in Q4 2026 to Q1 2027 with no first-party pricing we could verify. Committing frontier spend to Spark now means re-evaluating within the quarter. That double migration, not Spark's age, is why this says Wait.
Wait, if you buy at the frontier tier. Spark 1.3 is a solid GA model with 1M context, but Muse Wonder, Meta's most capable model, in preview since 12 October at frontier level, is expected to reach GA in Q4 2026–Q1 2027 with no first-party pricing we could verify. Evaluate Wonder now; commit frontier spend when GA pricing lands.
Overview
Muse is Meta's current flagship family, the line that superseded Llama inside Meta. Muse Spark 1.3 (2 September 2026, GA, 1M context, $1.25/$4.25) powers the Muse agent, and Muse Wonder (muse-wonder-1.0) entered preview on 12 October as Meta's most capable model.
What's new in this generation
Spark's cadence ran 1.0 (April 2026) and 1.2 (July), with the Glimmer 30B open-weights release in August, before Spark 1.3 in September. Wonder is the new frontier tier above Spark: real and in preview, only its GA date and pricing are rumored.
Features & capabilities
Wonder and Wonder Full are separate preview variants. Their current Intelligence Index v5 scores are shown in Benchmarks with reasoning settings and a check date; earlier-index launch scores should not be treated as the same scale. Spark 1.3 remains the GA workhorse with 1M context for agent workloads.
Pricing
Spark 1.3 lists at $1.25/$4.25 per million tokens. We could not verify a first-party Muse Wonder price. OpenRouter has listed muse-wonder at $4/$20 with a 1M-token context since 12 October, which would sit above Spark; treat that as indicative and wait for Meta's own number before production planning.
Moonshot's K3.1 Pro brings native visual understanding and a 1M-token context to Kimi products and the API. Highspeed is a faster, higher-priced variant.
Context window
1,048,576 tokens
Input / output per million
$2 / $10
Editorial recommendation
Rolling Out
Kimi's API homepage lists Pro at $2/$10 and Pro Highspeed at $4/$20 per million tokens, checked 1 Nov 2026. The weights promise does not yet establish a license.
K3.1 Pro is available in Kimi's products and API. Its downloadable weights are a separate, still-pending promise for 4 November; API availability should not be confused with an open-weight release.
Test the available API for long-context coding and knowledge work. Self-hosting decisions must wait for the promised weights and their actual license, due by 4 November according to Moonshot.
Overview
Moonshot announced K3.1 Pro on 23 October. It is available through Kimi.ai, mobile, Kimi Work, Kimi Code and the API; enterprise administrators control access in managed products.
What's new in this generation
K3.1 Pro follows July's K3. Moonshot describes a 3T-class model with native visual understanding; the Pro Highspeed variant offers the same capabilities and context at roughly two to three times the speed, according to the company.
Features & capabilities
Kimi's docs specify 1,048,576 tokens and low, high and max reasoning settings. Artificial Analysis lists K3.1 Pro at 46 on Intelligence Index v5, checked 1 November. Its vendor benchmark claims and independent scores are separate measurements.
Pricing
Kimi's own API homepage lists standard Pro at $2/$10 and Highspeed at $4/$20 per million input/output tokens. Moonshot promises full weights by 4 November; we have not verified a published license or downloadable K3.1 Pro weights.
Release history
2026-07-16: Kimi K3
What may come next
Kimi K3.1 Pro downloadable weights · By 4 Nov 2026 (company promise) · Official promise; weights pending
DeepSeek released the V4.1-Flash-1017 checkpoint on 17 October through its website, app, API and open-weight channels. It follows the original V4.1-Flash release on 10 September.
Context window
1M tokens
Input / output per million
$0.30 / $1.20
Editorial recommendation
Adopt Now
Peak-rate default $0.30/$1.20 per million input/output tokens; off-peak $0.15/$0.60. Peak windows and holiday rules are shown with the pricing sources.
V4.1-Flash shipped with MIT open weights on day one at roughly $0.15/$0.60. At this price, and with weights in hand, there is nothing to wait for.
Adopt now. V4.1-Flash shipped 10 September with MIT open weights on day one at roughly $0.15/$0.60. At this price, and with weights in hand, there is nothing to wait for, the next release is only a December cycle estimate.
Overview
DeepSeek is the aggressive end of open-weight pricing. The current API checkpoint is DeepSeek V4.1-Flash-1017 (17 October 2026, 1M context), released through DeepSeek's products, API and open-weight channels. It follows the original V4.1-Flash release on 10 September. The cycle ran V3.2 (December 2025), V4 (April 2026) and V4-Pro (13 August).
What's new in this generation
V4-Pro landed 13 August and V4.1-Flash followed 10 September with MIT weights on day one, no API-only window. Together they keep DeepSeek at the aggressive end of open-weight pricing through the densest flagship month of the year.
Features & capabilities
V4.1-Flash is not the absolute frontier, but nothing else is close on cost per point: it is the self-hosting and cost-sensitive-volume pick in our Benchmarks, and the current-generation reference against which other open-weight drops are measured.
Pricing
DeepSeek's pricing page lists deepseek-flash at $0.15 in / $0.60 out per million tokens off-peak and twice that at peak hours (01:00–04:00 and 06:00–10:00 UTC on weekdays); Frontrumors lists the off-peak rate. The 10 September weights are MIT-licensed, so self-hosting cost is your own infrastructure.
Release history
2025-12-01: DeepSeek V3.2
2026-04-24: DeepSeek V4
2026-08-13: DeepSeek V4-Pro
2026-09-10: DeepSeek V4.1-Flash
What may come next
DeepSeek V4.2 (cycle estimate) · Dec 2026 (estimated from cycle) · Low, cycle estimate only
Large 4 entered public preview on 6 October and its weights were released on 29 October. The model card lists 1.05T total / 52B active parameters and 1M context; the API's Public Preview label remains.
Context window
1M tokens
Input / output per million
$0.70 / $2.10
Editorial recommendation
Rolling Out
Mistral docs list 1M context, $0.70/$2.10 per million input/output tokens and $0.07 cached input; checked 2 Nov 2026. The released weights use the Mistral Ecosystem Open-Weight License, with separate commercial terms for Model-as-a-Service businesses.
Mistral released Large 4 weights on 29 October, and its public Hugging Face manifest and license were checked on 2 November. Its API docs still label the model Public Preview. Evaluate the available API or weights on your workload and review the separate commercial terms for Model-as-a-Service businesses.
Evaluate the API preview or the released weights. Public weights make self-hosting an available option, rather than an upcoming promise; infrastructure needs and the Mistral Ecosystem Open-Weight License still matter. Model-as-a-Service businesses need separate commercial terms from Mistral.
Overview
Large 4 entered public preview on 6 October and Mistral announced its weights release on 29 October. Its model card lists 1.05T total / 52B active parameters, a vision encoder and 1M context. Medium 3.5 remains the GA workhorse; the Large 4 API docs still say Public Preview.
What's new in this generation
The October weights commitment is fulfilled. On 2 November, we checked Mistral's public Hugging Face repository: its manifest lists 54 safetensors weight shards, plus configuration and license files. The small configuration file returned HTTP 200. We did not download or run the full model.
Features & capabilities
The released checkpoint accepts text and images and supports low, medium and high reasoning effort. Mistral positions it for coding and defensive security; vendor benchmark claims are separate from the dated independent measurements in our Benchmarks. Open weights are available under an ecosystem license, rather than plain MIT terms.
Pricing
Mistral's docs list $0.70 input, $0.07 cached input and $2.10 output per million tokens, checked 2 November. The Mistral Ecosystem Open-Weight License requires Model-as-a-Service businesses to obtain separate commercial terms; self-hosting also brings your own infrastructure costs.
Release history
2025-12-02: Mistral Large 3
2026-04-28: Mistral Medium 3.5
What may come next
Mistral Large 4 API general availability · No GA date announced; docs still say Public Preview · Weights released; API lifecycle timing unannounced
Qwen3.8-Max (3 Aug) tops Alibaba's line. The Max cadence has run roughly quarterly through 2026.
Context window
1M tokens
Input / output per million
$2 / $6
Editorial recommendation
Proven
Alibaba International (Singapore) / US (Virginia) API $2/$6 per million input/output tokens; Global-scope rates differ. Provider-specific regional and caching conditions apply.
Qwen3.8-Max (3 August) is fully rolled out, documented, stable-priced, and still Alibaba's top model with no dated successor, only a Nov to Dec cycle estimate. Late in the cycle is a feature here, not a warning. This is the battle-tested pick, especially for multilingual and Alibaba Cloud stacks.
Proven. Qwen3.8-Max has been available since 3 August, fully rolled out, and remains Alibaba's top model with no dated successor. Adopt it if it fits today, particularly for multilingual and Alibaba Cloud stacks, and re-evaluate next quarter; the Max line refreshes roughly quarterly.
Overview
Qwen is Alibaba's flagship family. Qwen3.8-Max (3 August 2026, GA, 1M context, $2/$6) tops the line. The Max cadence has run roughly quarterly: Qwen3-Max (September 2025), 3.5-Max (February 2026), 3.7-Max (May), 3.8-Max (August).
What's new in this generation
3.8-Max is a steady continuation of the quarterly Max rhythm rather than a clean-sheet jump, holding Alibaba's top tier through the autumn flagship season while Western labs shipped new generations in September and October.
Features & capabilities
Best for multilingual workloads and Alibaba Cloud stacks in our Benchmarks, a mid-cycle flagship to adopt on merit today. No successor has a published date; the Nov–Dec 2026 window is a cycle estimate only.
Pricing
$2/$6 per million tokens as reported by OpenRouter and model trackers, varying by region and provider.
Release history
2025-09-05: Qwen3-Max
2026-02-16: Qwen3.5-Max
2026-05-20: Qwen3.7-Max
What may come next
Next Qwen-Max (cycle estimate) · Nov–Dec 2026 (estimated from cycle) · Low, cycle estimate only
The open-weight line that defined 2025 has been superseded inside Meta by Muse. Behemoth remains unreleased.
Context window
10M tokens (Scout)
Input / output per million
Not verified / Not verified
Editorial recommendation
Legacy
Open weights; hosted pricing varies by provider.
Pricing last checked: not verified for Llama 4 Scout / Maverick
Meta's open-weights line effectively ends at Llama 4. Behemoth is indefinitely delayed and likely cancelled as a public release after the Muse pivot, and Meta's frontier effort now runs through Muse. Llama 4 still works, Scout's 10M-token context included, but there is no successor coming to plan around. Treat it as legacy: fine to run, wrong to build new plans on.
Legacy. Meta's Llama line effectively stops at Llama 4, and Llama 4 Behemoth is indefinitely delayed and likely cancelled as a public release after the Muse pivot. Build on Muse, or on Llama 4 as-is, Scout's 10M-token context included, but do not plan around Behemoth.
Overview
Llama is Meta's open-weight line, the family that defined 2025. Its current release is Llama 4 Scout / Maverick (5 April 2025, Limited), with Scout shipping a 10M-token context. The dated history here runs Llama 3.1 (July 2024) and 3.3 (December 2024) into Llama 4.
What's new in this generation
Nothing, and that is the story. The line that defined 2025 has been superseded inside Meta by Muse (Spark, Glimmer and now the Wonder preview). Llama 4 Behemoth remains unreleased.
Features & capabilities
Llama 4 as shipped is still usable on its own merits, headlined by Scout's 10M-token context. But it is a dead-end line: new investment belongs on Muse or another maintained family, per our reporting on the Wonder preview.
Pricing
Open weights; hosted pricing varies by provider. There is no first-party per-token list price to track.
Front